Title:	Advanced Digital Forensics with Focus on Deep Learning Foundation Models Research
DNr:	Berzelius-2025-423
Project Type:	LiU Berzelius
Principal Investigator:	Shizhen Chang <shizhen.chang@liu.se>
Affiliation:	Linköpings universitet
Duration:	2026-01-16 – 2026-08-01
Classification:	30501
Homepage:	https://wasp-sweden.org/people/shizhen-chang/
Keywords:

Abstract

Digital forensics is becoming increasingly reliant on advanced computational methodologies to keep pace with the evolving landscape of digital crime. The rise of deep learning and machine learning has significantly improved various aspects of forensic investigations, particularly in image and video analysis. However, the implementation of state-of-the-art deep learning models requires substantial computational resources, specifically GPU-based processing capabilities. This proposal seeks to secure GPU resources to develop and deploy sophisticated deep learning models for critical digital forensics tasks, including image manipulation detection, criminal activity analysis, and image restoration techniques like inpainting and watermark removal. This project seeks to address these challenges by leveraging diffusion models, a powerful class of generative models, to advance image manipulation detection task. Diffusion models, particularly Denoising Diffusion Probabilistic Models (DDPMs) and Latent Diffusion Models (LDMs), have shown remarkable success in generating high-quality, diverse visual data, making them ideal for simulating a wide range of manipulations. By applying these models, we aim to create a comprehensive, high-quality dataset of synthetically manipulated images that accurately represents real-world forgery techniques. This dataset will serve as a robust foundation for training and evaluating cutting-edge image manipulation detection algorithms. The project will be executed in three primary phases: dataset generation, algorithm development, and evaluation in real-world scenarios. In the first phase, diffusion models will be employed to generate a diverse set of manipulated images. These images will mimic various forgery techniques, such as splicing, removal, and copy-move, across a wide range of environments and manipulation types. The generated dataset will be annotated with detailed information on the types, locations, and nature of the manipulations, making it suitable for rigorous training and evaluation of detection algorithms. In the second phase, novel detection algorithms will be developed by integrating state-of-the-art deep learning techniques, such as convolutional neural networks (CNNs), transformers, and generative adversarial networks (GANs), with diffusion models. These algorithms will be designed to detect subtle image manipulations by learning the unique signatures left by different manipulation methods. Special emphasis will be placed on improving detection accuracy, robustness, and computational efficiency. In the third and final phase, the effectiveness of the developed algorithms will be tested and validated in real-world scenarios. This will include assessing their performance on low-quality images, compressed videos, and images manipulated using techniques not encountered during training. The evaluation will also involve collaboration with social media platforms, law enforcement agencies, and digital forensics teams to deploy and refine the algorithms in real-time environments.